Interaction Dynamics as a Reward Signal for LLMs

TL;DR

This paper proposes TRACE, a novel reward signal based on interaction dynamics in dialogue embeddings, which improves alignment and understanding of conversational success in large language models.

Contribution

Introduces TRACE, a geometric interaction-based reward signal for LLMs, demonstrating its effectiveness and complementarity to traditional textual analysis methods.

Findings

Interaction dynamics achieve 68.20% accuracy, comparable to transcript analysis at 70.04%.

Hybrid models combining dynamics and text reach 80.17% accuracy.

Interaction patterns serve as a privacy-preserving predictor of conversational success.

Abstract

The alignment of Large Language Models (LLMs) for multi-turn conversations typically relies on reward signals derived from the content of the text. This approach, however, overlooks a rich, complementary source of signal: the dynamics of the interaction itself. This paper introduces TRACE (Trajectory-based Reward for Agent Collaboration Estimation), a novel reward signal derived from the geometric properties of a dialogue's embedding trajectory--a concept we term 'conversational geometry'. Our central finding is that a reward model trained only on these structural signals achieves a pairwise accuracy (68.20%) comparable to a powerful LLM baseline that analyzes the full transcript (70.04%). Furthermore, a hybrid model combining interaction dynamics with textual analysis achieves the highest performance (80.17%), demonstrating their complementary nature. This work provides strong evidence that for interactive settings, how an agent communicates is as powerful a predictor of success as what it says, offering a new, privacy-preserving framework that not only aligns agents but also serves as a diagnostic tool for understanding the distinct interaction patterns that drive successful collaboration.